Author Affiliations
College of Mathematics and Informatics, South China Agricultural University/Guangdong Province Agricultural Data Engineering Research Center, Guangzhou 510642, Chinashow less
Fig. 1. Collecting thermal images
Fig. 2. Schematic diagram of Faster-RCNN
Fig. 3. Flow chart of segment
Fig. 4. Multi-channel grayscale
(a): Default grayscale; (b): R channel; (c): G channel; (b): B channel
Fig. 5. Otsu segment
(a): Thermal image of breeding eggs; (b): Default grayscale; (c): Otsu algorithm
Fig. 6. Structure of BP neural network
Fig. 7. Trend of total loss function
Fig. 8. Result of eggs detection
Fig. 9. Some segment of egg’s thermal images
(a): Thermal images; (b): Otsu algorithm; (c): BP neural network
Fig. 10. Segment’s images of egg and air cell
Fig. 11. Change of air cell size
Fig. 12. Correlation between the artificial measured value and thermal-image measured value
仪器 | 参数 | 参数值 |
---|
热像仪 | 测量范围 | -20~+120 ℃ | 测温精度 | ±0.05 ℃ | 图像分辨率 | 320×240 pixels | 波长范围 | 7.5~13 μm | 孵化箱 | 控温精度 | ±0.1 ℃ | 电热功率 | 300 W | 容蛋量(鸡蛋) | 1 232枚 |
|
Table 1. Instrument parameter value
初始学习速率 | 动量系数 | 平均精度均值/% |
---|
0.1 | - | 0 | 0.01 | - | 96.57 | 0.001 | - | 97.63 | 0.000 1 | - | 68.85 | | 0.5 | 99.65 | 0.001 | 0.9 | 99.74 | | 0.99 | 99.85 |
|
Table 2. Learning rate and momentum coefficient
隐藏层结构 | F1度量/% |
---|
1 000 | 81.45 | 1 000, 1 000 | 85.76 | 1 000, 1 000, 1 000 | 90.81 | 1 000, 2 000, 1 000 | 90.94 | 1 000, 3 000, 1 000 | 91.19 | 3 000, 3 000, 3 000 | 89.91 |
|
Table 3. Hidden layers’ optimization of BP neural network
初始学习速率 | 最大迭代次数 | F1度量/% |
---|
0.1 | | 69.76 | 0.01 | | 80.01 | 0.001 | 1 000 | 88.89 | 0.000 1 | | 90.38 | 0.000 01 | | 88.17 | 0.000 1 | 100 | 85.64 | 300 | 88.58 | 500 | 90.17 | 800 | 90.12 |
|
Table 4. Hyperparameters’ optimization of BP neural network
图像状况 | Otsu算法效果/% | BP神经网络效果/% |
---|
只存在一个蛋 | 68.86 | 87.17 | 有两个蛋存在 | 61.64 | 86.94 | 总体 | 65.25 | 87.02 |
|
Table 5. Comparison of algorithms for segment